L. Gao, G. Mittal, D. Zaretsky, D. Schonfeld, and P. Banerjee (USA)
Parallel processing, stream descriptor, stream architecture,FPGA.
In this paper we investigate the process and considerations for automatically generating streaming architectures from existing applications written for non streaming scalar processors. While the existing stream systems require the programs to be written in specific models, we develop a novel approach of identifying producer-consumer relationships from ordinary programs. As part of this approach, we use automatically generated stream descriptors along with operations such as subset analysis, dependence analysis, and stream concatenation to identify amenable data relationships. We demonstrate our results on a FPGA based platform. The automatically generated stream programs show significant performance improvements employing spatial and temporal data independence to increase large grain parallelism.
Important Links:
Go Back
